The familiar silhouette of bike and scooter delivery workers has become part of Paris’ urban landscape. For many city dwellers who rely on them to deliver meals to their door, these precarious workers remain largely “invisible” in surveys and public statistics.

Yet, the availability of quality data about online platforms’ delivery drivers is a major issue. Legally, the transposition into French law of the European Directive (EU) 2024/2831 on the legal framework around platform work (which aims to provide better protection to delivery couriers), expected before December 2 2026, makes it essential to have a better understanding of this population in order to shed light on regulatory choices.

Where occupational health is concerned, an expert appraisal by Anses (March 2025) exposes an alarming situation, and underlines the lack of available data for understanding the health status of these gig workers and implementing appropriate public policies.

It is in this context that France’s Santé-Course project was launched. Led by an interdisciplinary research team from the Institut de recherche pour le développement (IRD) and the Institut national d’études démographiques (INED), associations working with delivery people (Association de mobilisation et d’accompagnement des livreurs, AMAL; Collectif pour l’insertion et l’émancipation des livreurs, Ciel; Maison des livreurs de Bordeaux; Maison des couriers de Paris; Médecins du monde) and a peer group made up of couriers or former couriers, this project focused on documenting working conditions as well as delivery workers’ physical and mental health, based on a survey conducted among more than 1,000 of them in Paris and Bordeaux.

The study also examines exposure to occupational risks, police checks and cases of discrimination. Hereafter, we turn attention to the profiles of these workers and their working conditions, but the full results are available here.

What does platform work consist of ?

The rise of digital work platforms in France dates back about fifteen years and results from the conjunction of two series of factors: the adoption of new legal norms (notably the Novelli law of 2008 establishing self-employment status), on the one hand, and the generalisation of information and communication technologies as well as the democratisation of their use, on the other. The first point has gradually made the labour market more flexible and paved the way for massive employment of self-employed workers who are taken on by these platforms, while the second one has provided the latter with the conditions for their large-scale deployment.

À lire aussi : Requalifier ou réguler ? Les controverses du dialogue social des travailleurs des plates-formes

In the food delivery sector, digital platforms play a role as intermediaries between restaurant owners and customers, and between restaurateurs and deliverers. Their operations are based on matching algorithms, pricing, and disconnection that allow them to manage a vast network statutorily independent workforce, without having to resort to traditional company management methods.

Delivery drivers’ self-employed status places them outside the occupational health and safety regulatory framework that is applicable to employees. Their situation is similar to a return to task-based work, understood as project-based contractual work between clients and those who carry out the work.

As a result, social security contributions, which grant workers social protection and the legal obligations related to protecting workers, are transferred from the client to the self-employed worker. This puts delivery contractors in a highly precarious situation and makes them economically dependent on the platforms, which control access to deliveries and the terms of their remuneration.

A population that tends to be off the survey radar

Investigating platform delivery workers involves several methodological obstacles, the main one being admin-related: none of the company directories listing businesses located in France (Sirene, Sirus or Sine), usually used as sampling frames to draw samples from annual business surveys, allows reliable and exhaustive identification of platform deliverers. This makes it, therefore, difficult to know precisely their total number and their geographical distribution in France, thus making any approach by traditional sampling impossible.

Another problem is posed by the phenomenon of account leasing, which allows delivery drivers to carry out their activities under a third party’s account. This phenomenon also undermines the use of data from the platforms themselves, which lacks transparency (see the Anses March 2025 report).

As a result, only a direct canvassing protocol carried out in public places or community spaces is able to produce reliable data. This is how the Santé-Course project team managed : to meet delivery people at their pick-up destinations in Paris and Bordeaux.

The two French cities were selected because a significant part of these workers are concentrated there and they are home to partner associations of the Santé-Course project. In order to fully represent the diversity of situations experienced by delivery workers and, thus, obtain results that best reflect the reality of the entire population studied, an initial mapping survey of meeting points and the number of delivery people visiting them at different times of the day was carried out, which then served as a basis for the deployment of survey interviewers.

The survey was conducted during the first half of 2025, among delivery drivers over 18 years old, who had made at least one delivery via a digital platform in the month before the survey and were able to give informed consent. A total of 519 and 485 delivery people were interviewed in Paris and Bordeaux, respectively.

Nearly 1 in 2 delivery people spent an entire day without eating in the last twelve months

The results paint a remarkably homogeneous socio-demographic picture on several dimensions. The delivery workers are almost exclusively men (98.9%), immigrants (97.8%) and relatively young – their median age is 30 years old.

Their level of education, by contrast, is heterogeneous : while one quarter did not go beyond primary level education, nearly one in five went on to higher education, with significant differences between Paris (28.3%) and Bordeaux (9.6%).

Most of them recently arrived in France (median since 2020) and are mainly from West Africa and South Asia in Paris, from West Africa and North Africa in Bordeaux. Their administrative status is extremely fragile : nearly two thirds of them do not hold a residence permit.

This administrative hardship is coupled with material deprivation. The majority of the workers do not have a place to call their own : in Paris flat shares and lodging with people they know are the dominant trend, while communal supported housing and collective accommodation are more common in Bordeaux.

Even more worrying, nearly 18 % report living in unstable housing conditions (emergency accommodation, squats or welfare hotels). Food insecurity is just as significant : nearly one in two delivery people in Paris (48%) and more than one in three in Bordeaux (36.7%) report having spent at least a full day without eating, due to lack of money, over the past twelve months.

Nearly 73.5 % work under a third-party account

Those who were surveyed have been in business for some time: three quarters had never worked for a delivery platform before 2021, and more than one third of Parisian delivery workers started in 2024 or 2025.

Two platforms, Uber Eats and Deliveroo, largely dominate the market, but the simultaneous use of delivery services with several apps (or “multi-apping”) remains a very small minority, affecting less than 2% of them.

Economic dependence on this work is massive: 91% declare that delivery constitutes the bulk of their income, and about 95% do not engage in any other paid activity or are completing training alongside. Dependence on delivery work also appears to be largely constraine: nine out of ten deliverymen without a residence permit say they would cease or drastically reduce this type of work if they regularised their undocumented status.

Finally, the phenomenon of account rentals is massive: three quarters of delivery people work under the account of a third person, with a proportion reaching 81% in Paris. This phenomenon, which stems from the administrative precariousness of delivery people, many of whom are undocumented, considerably clouds the statistics produced by the platforms and highlights the need for surveys conducted directly with workers on the ground.

On average, 63-hour working weeks at a gross hourly rate of 5.83 euros

The delivery workers get an average gross monthly wage of 1,480 euros, or 880 euros after tax once all work-related charges are deducted. (including equipment and fuel expenses, insurance costs, taxes and, for three quarters of them, the rental cost of an account, which averages 528 euros per month and absorbs more than a third of gross income on its own).

The average gross hourly rate is a meagre 5.83 euros. This is well under France’s minimum wage (11.88 euros at the time of the survey) for significant volumes of work : on average 63 hours per week, six to seven days a week, ten months a year, with even more hours for those who rent an account. At this rate, they clock up 497 miles per month – such mileage is likely to be underestimated due to the omission of certain routes in the platform data.

This overview paints a picture of the “working poor”, a population forced to work extremely hard to earn an income after tax that remains well below the poverty line (set at 1,288 euros net per month for a single person).

The studies that will be conducted by our team over the coming months aim to shed light on the extent to which this situation affects the health of delivery workers.

More than half of the delivery drivers surveyed have already had at least one accident as a result of their work, and 44.8% of them believe that their health status has deteriorated compared to when they started working in the delivery industry.

This project received financing from l’Agence nationale de la recherche, l’Institut Convergences Migrations, la Ville de Paris, l’Inserm and l’Institut Paris Public Health at l’Université Paris Cité.

A weekly e-mail in English featuring expertise from scholars and researchers. It provides an introduction to the diversity of research coming out of the continent and considers some of the key issues facing European countries. Get the newsletter!

Les auteurs ne travaillent pas, ne conseillent pas, ne possèdent pas de parts, ne reçoivent pas de fonds d'une organisation qui pourrait tirer profit de cet article, et n'ont déclaré aucune autre affiliation que leur organisme de recherche.

131 sites and museums have suffered damage or been destroyed during the US and Israeli war against Iran between February 28 and April 8. Amid uncertainty over what will happen next as peace talks failed during the two-week conditional ceasefire, it is an opportune time to take stock of the state of Iranian cultural heritage.

With its vast territory and strategic position in West Asia, Iran has long been one of the principal centres of human activity and cultural development.

As one of the world’s oldest centres of civilisation, Iran has preserved an exceptionally rich archaeological and historical landscape shaped over several millennia. This heritage reflects a long sequence of cultural traditions, from the Palaeolithic (prehistoric) times through the Elamite kingdom (2700 BCE and 539 BCE), the Median dynasty (c. 700 to 550 BCE), the Achaemenid (559 to 330 BCE), the Parthian (247 BCE to 224 CE) and Sasanian (224 to 651 CE) empires and into the Islamic periods.

This continuity is visible in the country’s archaeological sites, historic cities, monuments, and museum collections. It is estimated that Iran contains several hundred thousand archaeological sites, although only a small proportion have been formally registered on the national heritage list since the beginnings of state heritage protection in the early twentieth century.

The international significance of this heritage is underscored by the inscription of 29 Iranian properties on the UNESCO World Heritage List, comprising 27 cultural and two natural sites. Last month, the UN cultural agency weighed in, voicing concern over the protection of Iran’s national treasures and sites of “cultural significance”, such as the Qajar-era Golestan Palace, following airstrikes. In a recent statement, the International Council on Monuments and Sites (ICOMOS) condemned any destruction – whether intentional or incidental – of cultural and natural heritage, deploring the “serious implications for cultural continuity” and the “risk of irreversible loss”, more broadly across the region as a result of the conflict.

What’s the damage?

An emerging official inventory of cultural damage recorded by the Ministry of Cultural Heritage, Tourism and Handicrafts of Iran shows that more than 131 archaeological sites, museums, and historical monuments (Figure 1.) in Iran have been damaged across 17 provinces and 26 cities.

The highest concentration of damage has been recorded in Tehran, where 61 sites were affected. It should be noted, however, that these figures are based on city-level assessment and do not include archaeological sites situated outside urban areas. In addition, historic urban fabrics are listed separately. The inventory recorded up to March 29 reveals a grave and highly uneven impact on Iran’s heritage, with destruction concentrated in some of the country’s most important historic and monument-rich cities.

The 1954 Hague Convention states that damage to any nation’s cultural property is a loss to the heritage of all humanity, which is why it requires international protection. Protecting cultural heritage is also tied to protecting human rights, including cultural rights, identity, memory, and human dignity.

As a result, intentional attacks on cultural heritage during armed conflict are not only morally unacceptable they could also violate international law and constitute war crimes, as confirmed by the International Criminal Court in the Al Mahdi case. This protection is further strengthened by UN Security Council Resolution 2347 of 2017, which emphasises the importance of safeguarding cultural heritage in conflict situations.

Tehran and Isfahan: the worst hit

What emerges from the inventory is not a scattered pattern of isolated incidents. It is a concentrated “geography of damage”, falling most heavily on provinces that hold some of Iran’s richest cultural assets, above all Tehran and Isfahan.

These are not marginal places in the historical map of Iran. They are among the country’s principal repositories of architectural memory, museum collections, 15th to 19th-century royal compounds, religious monuments, and civic heritage.

The most significantly damaged monuments in Tehran include Golestan Palace, Tehran’s Historic Arg, the Historic Grand Bazaar of Tehran, Marble Palace, the Historic Shahrbani Building, the Former Senate Building, Sepahsalar Mosque, and the Farahabad Palace Museum.

In Isfahan Province, damage has been reported at the Naqsh-e Jahan Square complex, the Chehel Sotoun Palace, the Abbasi Friday Mosque, etc.

The provincial distribution is among the most revealing aspects of the inventory. Tehran alone accounts for 61 counted sites, representing 46.6 percent of the total. Isfahan follows with 23 sites, or 17.6 percent. Together, these two provinces contain 84 damaged entries, equivalent to 64.1 percent of the inventory. When Khuzestan and Kurdistan are added, the top four provinces account for 108 sites, or 82.4 percent of all counted entries.

This means the damage pattern is not spatially even. It is clustered in provinces, where museums, palace complexes, historic neighbourhoods, old institutions or schools, and monumental architecture are densely concentrated.

The hypothesis of a strategically targeted assault

The typological profile of the damaged heritage is equally telling. The largest single group consists of historical houses, mansions, and residences totalling 33 entries. These are followed by civic and institutional buildings, such as schools, with 16 entries, and famous historical mosques, with 12.

The inventory also identifies historical forts, mills, and baths (hammam). The report additionally records 10 palaces or royal complex entries dating back to the 15th-19th centuries CE, indicating that the damage reaches deeply into architectural forms associated with old districts of the war-affected cities.

The document states that 50 museum units refer to museum components embedded within larger complexes, palace compounds, and multi-part heritage sites.

The cultural loss is therefore both architectural and institutional, affecting not only structures but also the curatorial and interpretive frameworks housed within them. According to Science, cultural institutions are taking measures to protect its moveable heritage, including boxing up museum items for safekeeping and installing the Blue Shield logo designed to indicate protected heritage on more than 100 cultural monuments.

The source also names 7 historic urban fabrics separately, suggesting that the true scope of impact extends beyond single monuments to wider urban heritage zones across Tehran, Isfahan, Sanandaj, Kermanshah, Qom, Khansar, and Tabriz. Old parts of the cities function as “layered cultural organisms”.

When an urban fabric is damaged, what is threatened is not only a set of buildings, but a continuity of streets, spatial memory, social practice, and architectural identity and art. Some of these fabrics were used for several hundred years and are a testimony of old traditions, artefact production, and Persian culture and identity.

If future surveys and analyses of Iranian sites are carried out, we will see that many sites outside the urban centres have been damaged. This damage has not been limited to buildings and museums, but has also affected archives of ancient manuscripts held in collections and places of worship such as mosques, churches and synagogues.

The bombing of the Cultural Heritage Office in Khorramabad city makes the deliberate nature of this destruction even clearer. These intentional acts of destruction are not limited to cultural heritage, but also extend to essential infrastructure, such as the unfinished Bridge in Karaj, the Pasteur Institute, and universities such as Shahid-Beheshti, Sharif and Elm-o Sanat (Science and technology).

What is dangerous here, as we see, is a portrait of cultural loss at multiple scales, from individual structures to complex heritage environments.

The chronological range of the damaged sites is striking. It extends from Kuh-e Khawaja in Sistan, one of south-eastern Iran’s most important archaeological sites, with remains dating to the Parthian, Sasanian, and early Islamic periods, to Siraf in Bushehr, the famous late antiquity and early medieval port city on the Persian Gulf, and to the tomb of Baba Taher in Hamedan, the celebrated 11th-century Persian poet.

The damage was not confined to historical monuments alone, but also reached the modern building of the Iranian Cultural Heritage office.

The targeting of cultural heritage in Iran, the historical memory and enduring identity of one of the world’s longest-lasting civilisations and an irreplaceable part of the heritage of humanity, was not incidental but systematic. Such acts must be condemned in the strongest possible terms.

They represent an assault not only on monuments, museums, and archaeological sites, but also on the cultural legacy, historical consciousness, and collective memory of humanity itself.

The right to remedy and the law on war reparations

Under international law, the law of reparations for war damage stipulates that a State responsible for an internationally wrongful act must make full reparation for any damage, whether material or moral.

This destruction must never be repeated. Urgent and immediate measures are now required to ensure the protection, documentation, stabilisation, and restoration of Iran’s damaged heritage.

These efforts must be undertaken without delay and supported at international level through coordinated action by cultural institutions, professional bodies, and relevant global organisations.

A weekly e-mail in English featuring expertise from scholars and researchers. It provides an introduction to the diversity of research coming out of the continent and considers some of the key issues facing European countries. Get the newsletter!

Parsa Ghasemi ne travaille pas, ne conseille pas, ne possède pas de parts, ne reçoit pas de fonds d'une organisation qui pourrait tirer profit de cet article, et n'a déclaré aucune autre affiliation que son organisme de recherche.

La polyvalence des modèles de fondation les transforme en une nouvelle « infrastructure » numérique, au même titre que le cloud ou Internet. Au lieu de reconstruire un modèle d’IA spécifique à chaque projet, on peut se brancher directement sur des briques généralistes existantes. C’est un des secrets qui permet de développer des applications si sophistiquées et qui restent accessibles aux non-spécialistes.

Les systèmes d’apprentissage automatique ne se limitent plus à des outils conçus pour une seule tâche, comme la traduction ou la recommandation de produits. Une transformation majeure tient à l’émergence des foundation models, ou modèles de fondation : de très grands modèles entraînés sur des volumes massifs de données pour acquérir des connaissances générales, réutilisables dans de nombreux contextes.

Dans les organisations, ils agissent comme un accélérateur potentiel de transformation, mais leurs effets sur le terrain obligent, pour l’instant, à nuancer les promesses spectaculaires. Comment fonctionnent-ils ? Comment sont-ils régulés et quels sont les obstacles à une adoption qui porte ses fruits ?

Comment fonctionnent les modèles de fondation ?

Les modèles de fondation reposent sur un principe simple : apprendre des structures générales à partir de très grandes quantités de données hétérogènes – textes, images, codes, sons, vidéos, bases de données ouvertes et contenus sous licence, ou une combinaison de ces types de données.

L’objectif de l’entraînement initial est de permettre au modèle d’identifier des régularités statistiques dans les données et de construire des représentations générales du langage, des images ou d’autres formes d’information. Sur le plan technique, ces systèmes utilisent le plus souvent des architectures de réseaux de neurones profonds. Durant la phase d’entraînement, le modèle apprend à prédire une partie manquante de l’information, par exemple le mot suivant dans une phrase ou une portion d’image, en ajustant progressivement des milliards de paramètres. Ce processus d’apprentissage, appelé pré-entraînement, nécessite des ressources de calcul considérables et constitue la base du caractère « généraliste » de ces modèles.

À lire aussi : Aux sources de l’IA : le prix Nobel de physique attribué aux pionniers des réseaux de neurones artificiels et de l’apprentissage machine

Une fois pré-entraîné, le modèle de fondation peut être adapté à des usages spécifiques, allant de l’analyse de sentiment ou la réponse à des questions jusqu’à des tâches plus techniques, comme l’assistance au diagnostic médical.

Cette adaptation peut se faire grâce à un ajustement supplémentaire appelé fine-tuning, par exemple en nourrissant un modèle généraliste d’imagerie médicale avec des radiographies spécifiques à une maladie permettant à l’outil d’apprendre à prédire l’évolution d’une pathologie précise.

Elle peut aussi passer par l’apprentissage avec retour humain (reinforcement learning with human feedback), qui consiste à faire évaluer plusieurs réponses par des humains pour inciter le modèle à privilégier des formulations claires et sécurisées plutôt que de simples suites de mots statistiques.

Enfin, cette adaptation peut s’opérer simplement par prompting, en guidant le modèle par des instructions textuelles du type : « Résume ce document en trois points. »

C’est cette capacité à être réutilisés dans de nombreux contextes qui explique pourquoi ces systèmes sont qualifiés de « modèles de fondation » : ils servent de base technologique à une large gamme d’applications. Par exemple, GPT-4 ou GPT-5 servent déjà de socle opérationnel à de nombreuses applications, à l’image de ChatGPT, tout en conservant un potentiel d’extension encore largement ouvert.

Comprendre l’écosystème : comment modèles de fondation, LLM et IA générative s’articulent-ils ?

Concrètement, les modèles de fondation ne sont pas une application en soi, mais une infrastructure de base. Ils marquent une évolution récente de l’intelligence artificielle (IA), rendue possible par la combinaison de trois facteurs : l’explosion des données, les progrès des capacités de calcul et l’apparition de nouvelles architectures d’apprentissage.

Un tournant majeur intervient en 2017 avec l’architecture des transformers. Cette innovation permet de mieux capter les relations dans les données (notamment le langage), et devient la base de modèles comme BERT ou GPT. Lorsqu’un modèle de fondation (une expression née officiellement en 2021) est spécialisé dans le traitement du langage, on parle alors de grand modèle de langage (LLM). Et c’est sur cette base que se développent aujourd’hui les usages les plus visibles : ceux de l’IA générative, capable de produire du texte, des images, des vidéos, du son ou du code – qui est devenue grand public et commercialement viable avec le lancement de ChatGPT, fin 2022, marquant le passage de l’infrastructure à l’usage de masse.

Le rapport entre ces modèles de fondation et l’IA générative peut être comparé à celui d’un « moteur » par rapport à sa « fonction ». Le modèle de fondation est ce moteur puissant, pré-entraîné sur des données colossales et conçu pour être adaptable à une multitude de tâches. L’IA générative, quant à elle, est la fonction d’application finale : c’est la capacité de ce moteur à produire un contenu inédit.

Concrètement, si l’on prend un modèle de fondation comme GPT-4 (le moteur), on peut l’utiliser pour analyser des milliers d’avis clients (une tâche purement analytique). Mais lorsqu’on lui demande de rédiger un e-mail, on active alors sa fonction d’IA générative. De la même manière, dans le domaine visuel, un modèle de fondation entraîné sur des millions d’images peut servir de moteur aussi bien pour détecter une anomalie sur une radiographie médicale (classification) que pour dessiner un paysage imaginaire à partir d’une simple phrase (IA générative).

Promesses d’efficience et réalités de terrain

Cette polyvalence signifie que les modèles de fondation tendent à devenir une nouvelle « infrastructure » numérique, au même titre que le cloud ou Internet : au lieu de reconstruire un modèle d’IA spécifique propre à chaque projet, les acteurs économiques se branchent directement sur ces briques généralistes existantes.

Dans les organisations, ils agissent comme un accélérateur potentiel de transformation, mais leurs effets sur le terrain obligent à nuancer les promesses de gains de productivité spectaculaires. Beaucoup d’entreprises peinent encore à dégager un retour sur investissement évident pour l’automatisation administrative, constatant souvent que les modèles de fondation ne réduisent pas la charge de travail, mais l’intensifie : les employés doivent désormais consacrer davantage d’énergie à vérifier et à corriger les résultats.

Par ailleurs, l’assistance aux experts (aide au code, à la décision) se heurte à une « frontière technologique en dents de scie » : face à une tâche donnée, le modèle peut exceller, mais s’avérer contre-productif s’il est utilisé aveuglément en dehors de sa zone de compétence.

Néanmoins, ces modèles permettent de créer de nouveaux services comme la personnalisation de la relation client à grande échelle. Mais pour libérer ce potentiel, la simple mutualisation technologique ne suffit pas. Il faut impérativement repenser l’organisation du travail en formant les employés pour leur donner l’autonomie nécessaire face à la machine.

Les débats européens sur la régulation

En Europe, les enjeux se sont cristallisés dans les discussions autour de l’AI Act, qui introduit une catégorie spécifique pour les « systèmes d’IA à usage général », dont les modèles de fondation sont l’exemple emblématique. L’idée est de ne plus réguler uniquement les cas d’usage finaux, mais aussi ces briques génériques qui irriguent tout l’écosystème.

La Commission nationale de l’informatique et des libertés (Cnil) s’est également saisie de ces enjeux à travers un plan d’action consacré à l’intelligence artificielle, visant à accompagner l’innovation tout en garantissant la protection des droits fondamentaux. Elle met notamment l’accent sur la protection des données utilisées pour entraîner les modèles, la transparence des systèmes ainsi que le développement d’IA respectueuses de la vie privée.

Dans ce contexte, l’entraînement de ces modèles soulève aussi des défis importants au regard du règlement général sur la protection des données (RGPD), notamment concernant l’origine des données utilisées, la possibilité pour les individus d’exercer leurs droits sur leurs données et la capacité technique des systèmes à supprimer ou de ne plus exploiter certaines informations après leur intégration dans l’apprentissage. Pour les entreprises, cela signifie que ces technologies doivent être intégrées dans des démarches structurées de conformité, de documentation et de gestion des risques.

La question devient donc : dans quelles conditions utiliser les modèles de fondation ? Cela implique une gouvernance claire entre fournisseurs, intégrateurs et utilisateurs, des exigences de transparence et de documentation, l’anticipation des impacts sur l’emploi à travers la formation et la reconversion ainsi qu’une articulation avec les politiques de responsabilité sociétale des entreprises (RSE), afin d’évaluer leurs effets sociaux, organisationnels et environnementaux.

Sabrine Mallek ne travaille pas, ne conseille pas, ne possède pas de parts, ne reçoit pas de fonds d'une organisation qui pourrait tirer profit de cet article, et n'a déclaré aucune autre affiliation que son organisme de recherche.